OpenAI has launched two cybersecurity products that put AI agents on both sides of the security fight. Daybreak Red Team attacks a web app to find its flaws, while Daybreak Defense Factory drafts the fixes. The pair was announced on 12 November 2026 at the Cyber at the Frontier keynote.
How Red Team works
Red Team sends a coordinated team of agents at a customer web application. The agents read source code and application context, work out how an attacker could get in, and confirm whether each weakness is real.
They also check findings from scanners, bug bounty reports and disclosure reports. Then they retest patches to see whether the hole is truly closed.
The testing runs in isolated environments, and a system OpenAI calls Guardian reviews what the agents may touch. Operators set the scope in advance, so the agents stay inside the agreed boundaries. Access is by application, and Red Team is in early access rather than general release.
How Defense Factory closes the loop
Defense Factory picks up where Red Team stops. It takes a finding, from Red Team or from a tool the team already uses, traces it through the code and prepares a patch. Engineers review the fix before it becomes a draft pull request, and nothing merges on its own.
The product ships as a plugin inside ChatGPT, in research preview, for workspaces that already have Daybreak access. Teams connect a GitHub repository, then run one-off scans, watch new commits or scan on a schedule. It also works with findings from Wiz, Snyk and Endor Labs.
The model doing the work
The Register reports that Defense Factory runs on GPT-6 Cyber, the security model OpenAI introduced on 5 November. OpenAI rates the model Critical for cybersecurity risk under its own Preparedness Framework, its highest tier, and restricts it to approved defenders.
Its system card puts intended exploit success at 48.45 percent on the ExploitGym benchmark, against 42.43 percent for GPT-6 Astra.
The same model recently validated 948 previously unknown privilege escalation flaws in the Linux kernel, in research OpenAI published on 4 and 5 November. That result is the kind of capability the two new products package for security teams.
OpenAI president Greg Brockman, who led the keynote, put the pitch this way in his announcement post. “We’re turning our lessons from defending OpenAI into practical tools for our customers.”
What OpenAI has not said yet
OpenAI has published no pricing for either product, and neither has a date for general availability. It has also published no independent results for the products themselves, and no figures for accuracy or false positives.
Early buyers are therefore taking the capability on trust, judging it on the model’s published benchmarks rather than on product results.
That gap matters because the tools touch live code and live systems. A missed flaw leaves a door open, and a wrong patch can break working software.
OpenAI’s answer is human review at every merge, plus the scope controls around Red Team. Whether that is enough will only show once customers run the tools on real codebases.
A security push that has been building all year
Daybreak itself is not new. OpenAI launched the program in May 2026 with partners including Cloudflare, Cisco, CrowdStrike, Oracle and Zscaler, and it has since committed $1 billion in subsidized access, training and support for frontline defenders.
The launch follows OpenAI securing $11 billion in the first close of a funding round at a $1.25 trillion valuation.
The last few days have been busy for the program. OpenAI open-sourced its Codex Security command line tool on 11 November, and Daybreak Red became available in Amazon Web Services GovCloud for approved government customers.
The two launches on 12 November package that work into products for security teams, one day after Google shipped Gemini 4 Hydrogen for coding and agents.
